Mentoring at the Global Youth AI Hackathon

Mentoring at a youth hackathon is mostly an exercise in scope reduction. Teams arrive with an idea sized for a company and a weekend in which to build it, and the single most useful intervention is usually a question about what can be cut.

The Global Youth AI Hackathon, organised by Tencent in Shenzhen in May 2026, brought together young builders working on agentic systems. I mentored across the event.

What the teams got stuck on

Almost none of the difficulty was model capability. The models were fine. The difficulty was in the parts around them, which is exactly where production agent systems get hard too, so the lesson transfers further than a hackathon deserves.

The recurring pattern was an agent given too much freedom too early. A loop with broad tool access and no stopping condition looks impressive for ten minutes and then burns an afternoon of budget retrying a call that was never going to succeed. Teams that constrained the action space first and widened it only when the narrow version worked finished with something demonstrable.

The second pattern was evaluation. Several teams had no way to tell whether a change had made their agent better, so every iteration was a matter of opinion. Twenty examples in a spreadsheet with an expected outcome beside each one is not sophisticated, and it is enough to convert argument into measurement.

What I tried to leave them with

Build the boring version first and make it work end to end, even if the intelligence in it is close to zero. A pipeline that reliably moves data from input to output with a stub in the middle can have a model dropped into it. A clever model with no pipeline around it is a screenshot.

Write down what the system is not allowed to do before you give it the ability to act. That constraint belongs in code beneath the model, not in a prompt asking the model to be careful.

And judge the result by whether somebody who was not in the room can use it. The gap between a demo that works when its author drives it and a system that works when a stranger does is the whole of engineering.

Why this is worth the day

The talent question in AI is frequently framed as a pipeline problem, as though the constraint were the number of people entering the field. In Shenzhen the constraint looks different. There is no shortage of capable young engineers. What is scarce is contact with people who have taken something to production and can say which of the hundred plausible next steps is the one that matters.

That is cheap to give and it compounds. A day of mentoring is a small thing set against the cost of a year spent optimising the wrong part of a system.